E3: Hybrid cache (SSM/conv) + unified hybrid support gate

[tlkahn]

Hybrid cache (SSM/conv) + unified hybrid support gate

Cache and admit-path foundations for lfm2 / nemotron_h / qwen3_5_moe linear-attn layers.

Spec: doc/plans/engine-inference-core.md § E3

Part of: #106 (E3) · Epic: #56

Tasks - cache

• SSM/conv state entries alongside KV entries (port SSMCacheEntry semantics)
• Prefill chunk boundaries respecting SSM checkpointing (nextChunkEnd)
• Create/reset/free on engine thread only

Tasks - support gate unification

Carried from PR #99 second review (finding 6), deferred until hybrid is first-class:

Today pure-dense qwen3_5 rejects hybrid via three overlapping checks:

1. linear_num_key_heads > 0 || linear_num_value_heads > 0
2. has_hybrid_layers (via shared reject_dense_common)
3. Loop over layer_is_global when has_explicit_layer_types

• Collapse into one shared hybrid gate helper
• Family-specific allow-lists for accepted hybrid shapes
• Remove pure-dense triple-reject once Stage E hybrid paths are real

Refs

• PR feat(engine): stage D6 - qwen3_5 dense full-attn deltas #99 review comment 5018410477 finding 6
• Issue Epic: Engine Inference Core Stage E - MoE + hybrid families #56 comment on unify hybrid/dense gate checks